Calibration

C-Eval → AA-GPQA Diamond

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap

AA-GPQA Diamond is estimated from C-Eval with a offset logistic curve fitted on 5 models measured on both: y = 0.8298 + (1.2000 − 0.8298) / (1 + exp(−200.00·(x − 0.9423))), R² = 0.67, cross-validated error 2.0 pp. It is used for 1 estimate.

Estimated model	C-Eval	AA-GPQA Diamond	Source
LongCat-Flash-Lite-Sparse	85.8%	83.0%	estimated ± 2.0 pp, low confidence